Log in

View Full Version : How consistant is CRF across profiles


benwaggoner
8th March 2010, 20:51
Running through some tests of different profiles targeting different devices, I realized I was assuming an answer to an unasked question:

To what degree can we expect the same CRF setting to provide the same visual experience with the same source, but with different profiles, or with other parameters different? I assume they wouldn't be identical (8x8 blocks aren't going to be pixel identical to 4x4), but how close is a reasoanble expectation?

Setting aside places where max bitrate/VBV would constrain things, should Baseline CRF 23 and High CRF 23 look pretty much the same, and thus the ratio of files sizes be said to be its compression efficiency advantage in that comparison?

When a VBV constraint kicks in, and if the same VBV is used for Baseline and High, we'd certainly expect a smaller file size gap, but the High encode to look better in places where Baseline is constrained by VBV.

Dark Shikari
8th March 2010, 20:59
There are two main factors that affect things:

1. With and without B-frames may not always be comparable (the way x264 handles internally with CRF is basically a wild guess).

2. Psy optimizations tend to increase bitrate at a given CRF.

benwaggoner
8th March 2010, 21:57
There are two main factors that affect things:

1. With and without B-frames may not always be comparable (the way x264 handles internally with CRF is basically a wild guess).
Ah, and so Baseline and Main/High could be different. Any idea how big a difference B-frames can have?

2. Psy optimizations tend to increase bitrate at a given CRF.
Right. I was imagining comparisons with the same psy optimizations, but differences in other parameters, like number of B/ref frames, Motion Search mode, etcetera.

Dark Shikari
8th March 2010, 22:04
Ah, and so Baseline and Main/High could be different. Any idea how big a difference B-frames can have?If we had a good formula for it, x264 would use something smarter than a wild guess ;)

You can try some testing.

nurbs
8th March 2010, 22:07
I was imagining comparisons with the same psy optimizations, but differences in other parameters, like number of B/ref frames, Motion Search mode, etcetera.
This might be helpful: http://doom10.org/index.php?topic=187.0